Hi, I’m Nathan
I’m a CS student at Lawrence Tech and I run a small web dev shop called Gamache Technology Solutions. I’m also lead developer at Perfect Afternoon. Most of my paid work is WordPress and WooCommerce, and outside of that I spend a lot of time on ML research and keeping my homelab alive.
Where I work
Perfect Afternoon. I joined the agency in June 2022 as a WordPress developer and became lead developer in March 2026. I work hybrid, full time in the summers and part time while school is in session. It’s client work: WooCommerce stores, corporate sites and franchise location pages. A project there usually has a project manager, who also runs QA, and a developer, and sometimes a designer or a marketer. On mine I own the development and the deployment, and our monthly audits keep watch on each site after launch. Across the agency that’s 260+ websites for 100+ clients on 10+ servers. A few of those projects are written up: Corrigan Companies, Terminal Supply Co and Cottage Inn. As lead developer I have focused on taking manual work off the team, most of all with PANoodle, the AI operations platform I invented and built.
Gamache Technology Solutions. My own company, an LLC I founded in February 2024. The main client so far is SJD Taxi, where I turned an unfinished replacement for their 2015 booking system into a working one, work I split between my shop and Perfect Afternoon. That one has a full case study. I also co-founded CatalogXpert in March 2026 as a venture between Gamache Technology Solutions and JDW Ventures. I wrote its original demo and first sales engine, and the specs its AI sales assistant platform was built from.
- Based in Michigan, USA
- Studying Computer Science (AI), Lawrence Tech
- Graduating December 2026
- Working at Perfect Afternoon
- Running Gamache Technology Solutions
Most of my hours
Web development
This is where I spend most of my hours. I build and maintain WooCommerce stores for real businesses, which means the work is less “pretty landing page” and more “why is the cart total wrong when someone adds a bundle with a variable product on sale.” A lot of it is debugging: tracking down a dead PayPal script that’s killing bundle JS, figuring out why UPS is returning weird shipping rates, fixing template overrides that haven’t been updated in four major WooCommerce versions.
The stuff I build from scratch tends to be:
- Custom plugins for specific client workflows (HubSpot sync, dynamic product filters, kit/bundle logic)
- ACF-powered Gutenberg blocks that editors can actually use without breaking the layout
- Performance work: caching, deferred CSS, font loading, cleaning up JS that’s blocking render
- Integrations with third-party services (shipping, inventory, CRM, payment)
I care a lot about leaving sites in better shape than I found them. Most WordPress codebases I inherit have accumulated years of “temporary” fixes, and part of the job is quietly undoing that without breaking anything in production.
Research
Machine learning
My senior project, which I finished in May 2026, was on violence anticipation in video: predicting risk before something happens rather than classifying it after the fact. During the capstone I worked with XD-Violence and UCF-Crime, transformer-based architectures and streaming evaluation metrics, since most benchmarks assume you have the whole clip, which defeats the point. In October 2026 I rebuilt the benchmark and the evaluation from scratch, with a baseline that sees no video at all, and the first results that beat it came out of that. More on the capstone.
For my Machine Learning class I co-wrote an unpublished paper on process and outcome reward modeling for reasoning in small language models, using a larger teacher model to train 1.5B-parameter reward models. Mostly it taught me how finicky reward modeling is and how much variance there is between runs. More on the paper.
For my Deep Learning class at Lawrence Tech in spring 2026 I built Detroit Pulse on my own. It turns public scanner radio for metro Detroit into a live incident map: Whisper transcribes the audio, a local Qwen2.5-7B model reads each transmission for its address and the kind of call, and a correlation engine groups transmissions into incidents. I also built a QLoRA pipeline for the address step, trained a LoRA adapter with it and ran the fine-tuned model in Detroit Pulse. More on Detroit Pulse.
What I actually enjoy is the full pipeline: dataset handling, training loops, evaluation design, and thinking about what it takes to run any of this in production with real latency constraints.
Systems
Homelab and systems
I run three servers in a rack at home, on two UPSs: a web server, a game server for friends and a NAS. Backups run every day. This website runs from the rack, and so does the vacation app I built for my family. It’s a low-stakes way to practice the infra stuff I don’t get to do much in WordPress land: Linux administration, networking, automation and keeping services up without babysitting them. More on the homelab.
Tech I use
What I use day to day.
Web
Machine learning
AI apps and data
Infrastructure
Next
What I’m working on improving
- Cleaner architectures without overengineering. It’s easy to reach for patterns that don’t pay off at the scale I actually work at.
- Getting ML models out of notebooks and into things that run fast and reliably in production.
- Systems fundamentals: networking, observability, and deployments that aren’t scary.